---
title: Local-Enhanced Guidance
url: https://www.emergentmind.com/topics/local-enhanced-guidance
type: topic
---

# Local-Enhanced Guidance

Searching arXiv for the cited papers to ground the article in current records.
Local-enhanced guidance denotes a family of mechanisms in which a planner, controller, or neural network augments its default update rule with fine-grained local cues while preserving a broader optimization, search, or inference process. In the literature, those cues appear as short spatiotemporal paths for agents, reference-path embeddings for navigation policies, channel-wise co-attention between local and global features, local planar priors in dense prediction, geometry-aware control terms in diffusion sampling, and guidance-modulated shrinkage in Bayesian regression. Across these settings, guidance is typically soft rather than mandatory: it biases local decisions, reweights features, or modulates update directions, with the stated aims of reducing congestion, improving robustness, sharpening local detail, or increasing signal–noise separation [2605.16855] [2603.13888] [2211.13919] [2603.11509] [2604.04964].

## 1. Conceptual scope and recurring design pattern

A recurring pattern is that local-enhanced guidance is introduced where an existing method already has a strong global backbone but weak short-range discrimination. In multi-agent pathfinding, the backbone is a configuration-based solver and the guidance is a short local path. In long-range navigation, the backbone is a recurrent policy and the guidance is a reference path embedding. In image enhancement and reconstruction, the backbone is a CNN-, transformer-, or state-space-based network and the guidance is a bidirectional exchange between local and global features. In diffusion sampling, the backbone is the reverse-time sampler and the guidance is a locally computed control update. In sparse regression, the backbone is a global–local shrinkage prior and the guidance is marginal association information embedded in the local variance map [2605.16855] [2603.13888] [2504.11375] [2603.11509] [2604.04964].

| Domain | Guidance carrier | Reported role |
|---|---|---|
| Lifelong MAPF | Guidance paths $\Phi[i]$ | Mitigate congestion, reduce waiting |
| Long-range navigation | Reference path embedding $e^{\rm path}$ | Soft heuristic for local execution |
| Image enhancement | Mutual guidance between branches | Bidirectional global-local exchange |
| CT correction | Global–local interaction guidance | Stripe extraction and correction |
| Diffusion sampling | Riemannian control update | Correct off-manifold drift |
| Sparse regression | $e^{\eta z_j^*}$ in variance mapping | Adaptive shrinkage |

This suggests a unifying interpretation: local-enhanced guidance is not a single algorithmic primitive but a design strategy for injecting high-value local structure into a system that would otherwise rely on coarser heuristics or globally symmetric updates.

## 2. Configuration-based pathfinding and lifelong MAPF

The most explicit formalization appears in configuration-based MAPF. In “Local Guidance for Configuration-Based Multi-Agent Pathfinding,” local guidance is recomputed around the current configuration $Q$ as a short spatiotemporal window of paths that anticipate future collisions over the next $w$ timesteps. Each agent’s cue is its first guided step $\Phi[i][1]$, and PIBT’s standard move ordering $\langle \dist(v,g_i),\epsilon\rangle$ is replaced by $\langle \mathbf{1}[\Phi[i][1]\ne v],\dist(v,g_i),\epsilon\rangle$. The paper states that LaCAM’s completeness is preserved because guidance only biases preferences and does not forbid exploration of collision-free neighbors; empirically it reports flowtime reductions up to $50\%$, with typical sweet-spots $w\in[10,20]$ and $\alpha\approx 3$ [2510.19072].

“Lifelong LaCAM with Local Guidance for Lifelong MAPF” lifts the same idea to the lifelong setting by combining a receding-horizon windowed framework with warm-started guidance. For agent $i$, the guidance path $\Phi[i]=(x_i^0,\dots,x_i^{w_\Phi})$ starts at $x_i^0=Q^t[i]$ and minimizes a cost that trades off remaining goal distance against predicted local collisions:
$$
\cost_i(\pi)
=
\dist\bigl(\pi[w_\Phi],g_i^t\bigr)
+
\sum_{\tau=0}^{w_\Phi-1}
\Bigl(1+\alpha\cdot\mathbf{1}\bigl[\chi(\pi[\tau],\pi[\tau+1])>0\bigr]\Bigr)
\cdot
\chi(\pi[\tau],\pi[\tau+1]).
$$
The resulting guidance is injected into PIBT through the lexicographic ranking
$$
\bigl\langle
\mathbf{1}\bigl[\Phi[i][1]\ne v\bigr],\;
\dist(v,g_i^t),\;
\mathrm{hindrance}_i(v),\;
\epsilon_i(v)
\bigr\rangle,
$$
so that following the guidance path is preferred, then true goal distance, then a hindrance estimate, then a small random bias. The lifelong solver warm-starts guidance by shifting the previous window ahead by one step, rebuilds it with small-horizon space–time A*, and carries the newly planned window forward after executing the first joint move [2605.16855].

The reported empirical effects are unusually concrete. In dense scenes, stop-counts per vertex drop by $60$–$80\%$, average wait-time drops by $20$–$30\%$, and throughput routinely surpasses fast suboptimal methods by $20$–$30\%$. On the random-64-64-10 map with $1{,}000$ agents, the reproduced comparison is:

| Method | Throughput | Runtime (s/step) |
|---|---:|---:|
| PIBT | 14.7 | 0.0015 |
| LaCAM (no LG) | 17.1 | 0.043 |
| Guided-PIBT | 16.0 | 0.015 |
| RHCR | 20.7 | 2.13 |
| LLLG | 21.1 | 0.21 |

The same study reports up to $+80\%$ throughput over RHCR in the densest cases, per-step overhead typically under $0.2\,\mathrm{s}$ for $n=1{,}000$ agents when $w_\Phi$ is small, and real-time viability up to $10{,}000$ agents with per-step planning remaining under $1\,\mathrm{s}$ [2605.16855].

## 3. Local guidance in navigation and reinforcement learning

In long-range embodied navigation, local-enhanced guidance appears as path conditioning rather than hard path tracking. “Path-conditioned Reinforcement Learning-based Local Planning for Long-Range Navigation” formulates the problem as a POMDP with observations
$$
o_t=\{I_t,\;p_t,\;o_t^{\rm prop},\;\tilde P_t\},
$$
where $\tilde P_t$ is a fixed-length egocentric reference path encoding. Each waypoint is converted to a $4$D feature, the path is processed by a self-attention block followed by a cross-attention layer with a learnable query, and the resulting embedding $e^{\rm path}\in\mathbb{R}^{128}$ is concatenated with the recurrent hidden state before the actor–critic heads. The reward combines task, regularization, penalty, and shortcut terms, but the paper explicitly notes that there is no explicit “stay on path” term; path-following emerges implicitly from conditioning plus the goal-driven reward. On the reported benchmarks, the baseline without path gives $\mathrm{SR}=0.832$ and $\mathrm{SPL}=0.746$, the method with an optimal path gives $\mathrm{SR}=0.8658$ and $\mathrm{SPL}=0.8165$, and the degraded-path condition falls back to approximately baseline performance with $\mathrm{SR}=0.8276$ and $\mathrm{SPL}=0.7427$ [2603.13888].

A more general control-theoretic version appears in “Enhancing Reinforcement Learning Agents with Local Guides.” There, a local guide is a policy $\pi_g:\mathcal{S}\to\Delta(\mathcal{A})$ paired with a confidence map $\lambda:\mathcal{S}\to[0,1]$. The paper’s Perturbed Action Guided policy-switching scheme uses the guide action when $\lambda(s)\ge \lambda^{-}$ but adds a bounded perturbation around it:
$$
\pi^k_{\rm PAG}(a\mid s)
=
\begin{cases}
a_g(s)+\beta^k\tilde a, & \tilde a\sim\xi_{\phi^k}(\cdot\mid s,a_g(s),\Phi),\ \text{if }\lambda(s)\ge\lambda^{-},\\
\pi^k_\theta(a\mid s), & \text{otherwise.}
\end{cases}
$$
A key technical point is that the same hybrid policy is used in Approximate Policy Evaluation, avoiding distribution-shift bias in the bootstrap target. Across six environments, the reported mean normalized AUC rises to $0.80$ for PAG, versus $0.57$ for Strict Action Guided, $0.43$ for Policy-Improvement Guided, and $0.32$ for unguided SAC [2402.13930].

A lighter-weight instance is the Destination Assistant Agent in “The User-Centric Geo-Experience.” It does not run a local graph planner; instead it continuously computes a great-circle bearing $\theta$ and a user-relative direction $(\theta-\alpha)\bmod 360^\circ$, then renders textual or AR cues at approximately $1\,\mathrm{Hz}$. The paper reports a $34\%$ reduction in critical navigation errors and a $93\%$ success rate for the last-$100$-meter task, explicitly characterizing the system as “which way” guidance rather than exact micro-routing [2507.06993].

## 4. Bidirectional global-local guidance in vision and imaging

In image enhancement, the phrase is often realized through mutual or interaction guidance between global and local branches. “Mutual Guidance and Residual Integration for Image Enhancement” introduces a two-branch network: a transformer-based global branch compressing image-wide context to a vector $\mathbf f_g$, and a CNN-based local branch retaining full-resolution feature maps $\mathbf F_l^t$. At each interaction stage, pooled local features attend to the global vector and the global vector reweights local channels via a compact channel-wise co-attention. Afterward, a residual integration scheme decomposes a global residual map into coarse-to-fine slices before adaptive reconstruction. On MIT-Adobe-5K-UPE, naive channel concatenation yields PSNR/SSIM $23.63/0.902$, one-way Local$\rightarrow$Global yields $24.98/0.923$, one-way Global$\rightarrow$Local yields $24.85/0.921$, and full bidirectional mutual guidance reaches $25.59/0.930$; increasing the number of guidance blocks from $T=1$ to $T=5$ improves PSNR from $23.50$ to $25.59$ dB [2211.13919].

In CT artifact correction, “Ring Artifacts Correction Based on Global-Local Features Interaction Guidance in the Projection Domain” explicitly decomposes a corrupted projection $P'$ by DWT into a low-frequency sub-band $P^{LL}$ and three high-frequency detail sub-bands. A VSS-based global branch corrects the smooth projection component, a Dense-block local branch models stripe artifacts, and the Global–Local Features Interaction Guidance module performs cross-attention in both directions before adaptive re-fusion. The reported Mayo’16 projection metrics improve from PSNR $34.59$ dB, SSIM $0.6867$, RMSE $0.0220$ at input to PSNR $58.85$ dB, SSIM $0.9997$, RMSE $0.0014$ after correction; on reconstructions the values move from PSNR $21.85$ dB, SSIM $0.3024$, RMSE $200.35$ HU to PSNR $50.84$ dB, SSIM $0.9977$, RMSE $7.19$ HU. An ablation without GLFIG drops projection PSNR by $3.3$ dB and increases RMSE by $43\%$ [2504.11375].

“Adaptive Low Light Enhancement via Joint Global-Local Illumination Adjustment” uses a different coupling. The Local Contrast Enhancement Network processes non-overlapping patches through enhancement and feature-modulation blocks, while a Local Discriminative Module predicts a readiness score $s_p$ and triggers early exit for sufficiently illuminated patches. The Global Illumination Guidance Network computes image-wide features $F_g$ and injects them into local patch processing through a Spatial Feature Transformation. On a $2{,}500$-image test set from LOL, MIT-Adobe FiveK, and SICE, the method reports PSNR $21.37$ dB and SSIM $0.8129$, exceeding the stated second-best values of $21.03$ and $0.8104$; on no-reference evaluation it reports LOE $351.92$ and NIQE $3.41$ [2504.00400].

Dense prediction offers another interpretation. “From Big to Small: Multi-Scale Local Planar Guidance for Monocular Depth Estimation” inserts Local Planar Guidance layers at resolutions $H/8$, $H/4$, and $H/2$, each predicting a plane for a corresponding $k\times k$ full-resolution patch. The ray–plane intersection formula
$$
d_i=-\frac{n_4}{n_1u_i+n_2v_i+n_3}
$$
converts those plane parameters into local depth estimates. On NYU v2 with a ResNet-101 backbone, $\delta<1.25$ improves from $0.845$ with ASPP and upconvolution to $0.863$ after adding LPG, and the full system reaches $0.871$; with DenseNet-161, the paper reports $\delta<1.25=0.885$ [1907.10326].

A related unfolding-based instance appears in “Local-Global Transformer Enhanced Unfolding Network for Pan-sharpening,” where PAN guidance is injected in the data module of each PGD stage, while the prior module uses a Local-Global Transformer that combines windowed self-attention and global Fourier-domain mixing. The reported implementation uses $K=2$ unfolding stages and a simple $\ell_1$ reconstruction loss [2304.14612].

## 5. Guidance in generative models and exemplar-based transfer

Example-based stylization provides an early and explicit notion of local guidance. “StyleBlit” distinguishes global guidance from local guidance by the spatial variation of the guiding channels, with normals, UV coordinates, and dense displacement fields given as representative local signals. For each target pixel $p$, the core mapping is
$$
u^*=\arg\min_{u\in\Omega_s}\|G_t(p)-G_s(u)\|,
$$
followed by chunk growth under a guidance error threshold. The final seed-hierarchy algorithm achieves $O(N)$ complexity, with reported performance of about $0.1\,\mathrm{s}$ for a $1$ megapixel frame on a single CPU core, about $0.01\,\mathrm{s}$ for a $4$K frame on a GeForce GTX 970, about $15$ fps on a Samsung Galaxy A3, and a speed-up on the order of $10^3\times$ over StyLit [1807.03249].

In video diffusion, local guidance is fused with global context. “Video Diffusion Models with Local-Global Context Guidance” conditions each fragment on a small set of local frames $y^i$ and positional masks $m^i$, while a sequential encoder produces a global embedding $z^i$ from the previous fragment. Local information is concatenated channel-wise, and the global embedding is injected through cross-attention inside a 2D-convolutional UNet with residual and self-attention blocks. The paper further introduces a two-stage training strategy in which the second stage uses the model’s own predicted fragment as the condition for the next ground-truth fragment. On Cityscapes, the reported comparison improves FVD from $141.3$ to $124.6$, LPIPS from $0.112$ to $0.069$, and SSIM from $0.690$ to $0.732$; on BAIR, PSNR improves from approximately $17.7$ to approximately $20.4$ and SSIM from approximately $0.797$ to approximately $0.863$ [2306.02562].

“Manifold-Optimal Guidance” reframes diffusion guidance as a local optimal-control problem on a noisy data manifold. Standard CFG uses
$$
s_{\rm CFG}=s_0+w\,\Delta s,
$$
where $\Delta s=s_c-s_0$, and the paper attributes oversaturation, texture artifacts, and structural collapse at large guidance scales to Euclidean extrapolation off the data manifold. MOG instead solves
$$
u_t^\star=-\beta\,M_t^{-1}\,\nabla_x\E(x_t,c),
$$
yielding the guided score
$$
s_{\rm MOG}(x_t)=s_0(x_t)+\beta(t)\,M_t^{-1}\,\Delta s(x_t).
$$
With a rank-$1$ anisotropic metric based on the unconditional score direction, the update can be applied in $O(d)$ by Sherman–Morrison, and the paper states that this yields virtually no added computational overhead. Auto-MOG further sets
$$
\beta_{\rm auto}(t)=
\gamma\,
\sqrt{\frac{s_0^\top M_t s_0}{\Delta s^\top M_t^{-1}\Delta s}},
$$
removing the need for manual guidance-scale tuning [2603.11509].

A task-specific dense-prediction instance is “SLENet,” where the Localization Guidance Branch produces a semantic map $M$ that modulates each decoder scale through affine parameters. On DeepCamo-test, the full model improves $F^w_\beta$ from $0.764$ in the unguided setting to $0.800$, while reducing MAE from $0.026$ to $0.022$ [2509.03786].

## 6. Statistical and algorithmic generalizations

Local-enhanced guidance is not confined to spatial fields. “Bayesian Global-Local Shrinkage with Univariate Guidance for Ultra-High-Dimensional Regression” embeds marginal association information directly into the nonlinear variance structure of a regularized horseshoe prior. The effective variance is
$$
\tilde\kappa_j^2=\frac{c^2A_j}{c^2+A_j},
\qquad
A_j=\tau^2\lambda_j^2 e^{\eta z_j^*},
$$
so high-guidance predictors require smaller $\lambda_j$ to reach the slab regime, while weakly supported predictors remain strongly shrunk. The paper proves prior concentration, posterior contraction, and guidance-induced shrinkage separation, and also proves robustness under uninformative guidance. Its BUGS-Active variant restricts local updates to an active set $A_n$, reducing per-iteration local-update complexity from $O(p)$ to $O(|A_n|)$; the reported scale reaches $p=1{,}000{,}000$, and in a DNA methylation study with $n=1051$ and approximately $850{,}000$ CpG sites, BUGS-Active improves test RMSE from $6.210$ for the unguided method to $4.882$ [2604.04964].

A closely related, though polarity-reversed, formulation appears in SMT-based model checking. “Global Guidance for Local Generalization in Model Checking” argues that IC3-style rules are local and can fail through myopic generalization, excessive generalization, or being stuck in a rut. Govind et al. add three rules—Subsume, Concretize, and Conjecture—so that the current set of frames acts as a tentative global proof guiding subsequent local reasoning. The paper reports that GSpacer is significantly more effective than both SPACER and sole global reasoning, and that it is insensitive to interpolation [2005.13301].

Several misconceptions recur across the literature. Guidance is often treated as if it were a hard constraint, but many of the strongest results arise when guidance remains soft: LaCAM preserves completeness because guidance only biases preferences; path-conditioned RL has no explicit path-following reward; and StyleBlit uses guidance to identify transferable chunks rather than to solve a global optimization [2510.19072] [2603.13888] [1807.03249]. Another misconception is that stronger guidance is always better. High CFG scales can induce oversaturation and structural collapse, and MAPF guidance parameters that are too large increase planning time or make agents overly conservative; this is why adaptive schedules such as Auto-MOG and empirically tuned windows such as $w\in[10,20]$ recur in practice [2603.11509] [2510.19072].

Taken together, these works define local-enhanced guidance as a broad methodological family: local cues are made explicit, encoded, and reinjected into an existing decision loop so that global objectives can be pursued with finer situational awareness. The implementations differ sharply across graph search, continuous control, inverse problems, generative modeling, and Bayesian inference, but the underlying logic is consistent: local information becomes most useful when it is informative enough to alter short-range behavior and lightweight enough not to destroy the scalability of the surrounding system.

Source: https://www.emergentmind.com/topics/local-enhanced-guidance